SANA-Video 2.0: Hybrid Linear Attention with Attention Residuals for Efficient Video Generation
Junsong Chen, Jincheng Yu, Yitong Li, Shuchen Xue et al. · arXiv · 2026
Abstract
A video diffusion transformer built to run at 720p on one GPU by mixing attention types rather than committing to a single one. Linear attention handles most of the token mixing at cost growing linearly with sequence length, while periodic softmax layers restore the full-rank interactions pure linear attention gives up. The hybrid is trained from scratch rather than converted from a pretrained network, and produces a 480p clip in about thirteen seconds on one H100.
Why it matters
The clearest recent evidence that linear attention works as one half of a hybrid rather than as a wholesale replacement — the same trade-off now shaping long-context language models.
https://arxiv.org/abs/2607.21553